EverywhereSafety/SEAD-SFT-v1

SEAD SFT v1

Supervised fine-tuning data for SEAD: A State-Based Perspective on Attack and Defense in Tool-Using Agents. The dataset trains SAGE, SEAD's state-aware defender, to decide whether a pending tool action should pass or block from the visible interaction history and candidate action.

Paper · Model weights · GitHub code · Project website · Everywhere Safety

Dataset overview

The release contains one action-level training example per JSONL row. Each example presents the defender with an ordered agentic conversation containing the safety policy, visible trajectory evidence, the pending candidate action, and the expected structured defender response.

Split	Examples	pass	block
Train	7,407	6,792	615
Validation	131	100	31

Training and validation use disjoint source trajectories. manifest.json records the release schema and split statistics.

Important fields

Key	Meaning
sample_id	Source-sample identifier used for grouped split separation; multiple accepted traces may share it within one split.
split	Dataset split: train or validation.
label	Gold safety decision: pass or block.
defender_details.agentic_messages	Ordered system, user, and assistant messages used directly for supervised fine-tuning. Each message contains role and content.

The central training field is defender_details.agentic_messages. It preserves the complete model-facing conversation rather than flattening the interaction into a single prompt. The final assistant message is a JSON action such as a direct submit decision or an intermediate read-only investigation step, allowing SAGE to learn both safety decisions and bounded state investigation.

Load with Datasets

from datasets import load_dataset

dataset = load_dataset("EverywhereSafety/SEAD-SFT-v1")
train = dataset["train"]
validation = dataset["validation"]

messages = train[0]["defender_details"]["agentic_messages"]
label = train[0]["label"]

Training contract

The public SEAD training code validates sample identity, split membership, the gold label, final submitted decision, message ordering, and structured assistant actions before tokenization.

SAGE is trained with the Qwen chat template, assistant-only loss, BF16 parameters, thinking disabled, and a 65,536-token sequence budget. At inference time, the working configuration uses ordinary generation with prompt-level JSON instructions and parses the final complete JSON object. Provider-level JSON mode or constrained decoding is not required.

Data provenance

The examples are derived from stateful tool-use trajectories built with OpenAgentSafety and related benchmark infrastructure. The source trajectories are balanced at the trajectory level between benign and harmful cases before creating action-level demonstrations. Harmful trajectories contribute both pre-closure pass decisions and harm-enabling block decisions, while benign trajectories teach utility-preserving pass behavior.

The data is released for research use subject to the applicable upstream dataset and benchmark terms. This card does not grant rights beyond those source terms.

Citation

@misc{shen2026seadstatebasedperspectiveattack,
  title={SEAD: A State-Based Perspective on Attack and Defense in Tool-Using Agents},
  author={Xinjie Shen and Junran Wang and Rongzhe Wei and Pan Li},
  year={2026},
  eprint={2609.34518},
  archivePrefix={arXiv},
  primaryClass={cs.CR},
  url={https://arxiv.org/abs/2609.34518}
}